Re: iSCSI: Markers

[Stephen Bailey <steph@cs.uchicago.edu>]

> If we decided it was the right technique for iSCSI, couldn't we just
> lift the content and put it in the iSCSI draft?

From a standardization standpoint, it's like JohnH said.  iSCSI is not
chartered to do anything that modifies the transport.  Even if it were
(which it isn't) the `experimental stigma' of TUF would certainly
spread to iSCSI if iSCSI ingested TUF in toto.

The allure of COWS is that iSCSI CAN specify the in-stream format
without saying anything about TCP segmentation---COWS in-stream format
solves the same problem as markers.  If iSCSI does that, it's a tiny
step, for implementations that can take it, to add TUF-compliant TCP
segmentation & PDU containment processing.

What would be bad is if the `performance community' (RDMA, iSCSI &
whomever else) decided that COWS is the best general solution to this
problem, and iSCSI chose a variant of COWS that didn't work for other
applications (current & future).

Basically, the choices, as I see them are:

  o nothing, FIM, COWS for unmodified senders
  o key/length or COWS for TUF-compliant senders

The key questions to make the choices come down to:

  1) Is PDU containment `worth it'?
  2) Is resynching in the absence of PDU containment `worth it'?
  3) Is 0-touch send for software clients `worth it'? 

Most seem to strongly believe that PDU containment is worth it but
there are some experienced, noteworthy objectors (e.g. Jim Williams).
If you don't believe PDU containment is sufficiently beneficial, then
you simply don't bother with TUF-compliant senders.

I have no idea what `most' think for 2) & 3), but I can speak for
myself.

I don't think resynching in the absence of PDU containment is worth
it.  That implies that I'd take nothing over FIM.  It also implies
that I'd take nothing over COWS w/o a TUF sender, except that if the
TUF sender approach uses COWS, having COWS in-stream is `free'.

I'm undecided on whether 0-touch send for software clients is `worth
it'.  If yes, I'd take key/length for a TUF sender.  If no,
I'd take COWS.

So the only free variable in my position is whether 0-touch software
senders are worth it.  I don't have the data (I'm sure you're
wondering why that's stopping me on this point, and not the others :^)
If yes, it's `nothing + key/length'.  If no, it's `COWS + COWS'.

Steph